I build ML systems for science and production: LLM agents with measured evaluation, weather forecasting models, genomics NLP.
Not from a CS background. Taught myself ML, shipped code upstream, earned PyTorch org recognition. Most of what I know came from having patches reviewed by maintainers who did not have to be kind about it.
B.Tech @ NIT Rourkela '27 | CGPA 8.02 | Odisha, India
|
|
|
|
-3_merged-FF6600?style=flat-square&logo=kubernetes&logoColor=white){kind=icon}
- graph_weather #3 / #238 — adaptive meshing for regional NWP. Currently running a controlled ablation on whether the mesh message-passing earns its parameters against a graph-free baseline; publishing the numbers either way.
- krkn-chaos/krkn-ai #389 — Elasticsearch config for composite chaos scenarios
- malariagen-data-python #1310 — haplotype network mixin
| Project | What it does | Highlights |
|---|---|---|
| Climate-Risk Analyst Agent — live | LangGraph agent returning typed, cited climate-risk reports — or refusing when the evidence is thin | 105-question SHA-frozen held-out set: recall@3 87%, citation validity 94%, zero false answers. 238 tests, ~$0.001/report, deployed on Cloud Run |
| Incident Evidence Compiler | Root-cause investigation service where the model proposes and deterministic code decides | RCAEval RE2-OB: top-1 0.932, MRR 0.959. Content-addressed evidence ledger, Postgres SKIP LOCKED worker queue, 304 hermetic tests, mypy strict |
| Weather Transformer | Physics-aware Transformer for 6-hour weather forecasting on ERA5 | Written from scratch in PyTorch — no nn.MultiheadAttention. 74 tests, beats persistence by 27% RMSE on a 1,316-sample held-out set |
| Fairness-Aware Credit Risk | Four bias-mitigation interventions measured against a tuned baseline under identical conditions, on two credit datasets | No intervention helped — and the reasons differ per dataset. Bootstrap intervals on every metric, 212 tests, SHAP adverse-action reason codes, model card. Includes a published self-audit of three defects in my own earlier version, including a proxy leak where dropping gender did not remove sex from the model. Every published number regenerated from an artifact, enforced by a test |
| Complaint Intelligence — live | RAG and retrieval benchmark over CFPB consumer complaints | 200K CFPB complaints processed end to end. MiniLM vs BGE, vector vs BM25 vs hybrid vs rerank, KMeans vs BERTopic — every component chosen on measured results, not defaults |
| LLaMA Task Agent | LoRA fine-tuned LLaMA-3.1-8B for agentic tool execution | Generates valid function calls with type-safe args from natural language |
| Bio Publication Analyzer | SciBERT + TF-IDF classifier for genomics articles | 81 tests, Europe PMC pipeline end to end |
